Evidence and its limits.

Pytxo publishes one performance figure. This page states how it was produced, what the corpus was, and which conclusions it does not support, so you can decide whether it transfers to your repository.

The figure

Weighted reduction
82.9%
Median per file
86.3%
Files in corpus
185
Raw fallbacks
0

1,253,675 bytes of source were reduced to 213,847 bytes of structural scaffold. Measured 2026-07-30 at commit 3acdc77 on a single Windows host.

The corpus

Tracked production source under crates, services, packages, Desktop and Web source, and tooling; supported Signal grammars only; tests, fixtures, stories, specs, vendor, dependencies, and generated output excluded. Files under 1,024 bytes were excluded.

Rust1251,010,931121,24588.0%
TypeScript54221,20391,36458.7%
JavaScript621,5411,23894.3%

What it does not claim

  • Not measuredNot a model-token saving. Tokenizer output was not measured.
  • Not measuredNot a cost saving. No provider bill was compared.
  • Not measuredNot a task-success claim. Agent outcome quality was not measured.
  • Not measuredNot a general-repository result. One repository, one commit, one host.

Caveats that weaken the number

  • The measured worktree was dirty at capture time, so the corpus reflects working state rather than a clean checkout of the recorded commit.
  • Files below 1 KiB and languages without a Signal grammar are excluded, which raises the weighted figure relative to a whole-repository read.
  • TypeScript reduced least in this corpus at 58.7%; Rust reduced most at 88.0%. A repository weighted toward TypeScript should expect a lower number.

Reproduce it

The raw result, including the per-file table for every one of the 185 files, is committed in the repository.